Blog
Ideas on organisational intelligence, AI-assisted engineering, and how the world is evolving.
The organisations that survive AI acceleration won't be the ones with the best agents. They'll be the ones that learned to govern them.
MCP servers, CLI tools, skills files, constraints - they all cost tokens. The question isn't how to minimize that cost. It's what your tokens are buying you.
Your agents aren't broken. They're amnesiac. And you've gotten used to it.
Two papers measured what happens when agents drift from their instructions. Both proposed better enforcement. Neither asked whether the rules were worth enforcing.
Engineering organizations are running experiments to figure out what organizing means when agents do the work. The interesting part isn't the experiments - it's where they keep hitting walls.
The industry is building defences against unauthorised prompt injection. The other direction - detecting when authorised, trusted instructions are steering agents wrong - gets handled manually. That works, up to a point.
Everyone's debating what replaces code review as an approval gate. But code review was also the last informal pipeline for tribal knowledge - and that's the part nobody's replacing.
